Replace reduce by mapreduce for kernelproduct

[theogf]

Typically

    return mapreduce(hadamard, kernels(κ)) do kernel
        kernelmatrix(kernel, X; obsdim = obsdim)
    end

should be more efficient since as the documentation says

mapreduce is functionally equivalent to calling reduce(op, map(f, itr); init=init), but will in general execute faster since no intermediate collection needs to be created.

Originally posted by @devmotion in https://github.com/_render_node/MDIzOlB1bGxSZXF1ZXN0UmV2aWV3VGhyZWFkMjU3MDk2NjA5OnYy/pull_request_review_threads/discussion

[theogf]

Running

using BenchmarkTools, KernelFunctions

k = KernelProduct(fill(SqExponentialKernel() , 100));
hadamard(x, y) = x .* y
function km1(κ, x)
     mapreduce(hadamard, κ.kernels) do k
            kernelmatrix(k, x)
        end
end
function km2(κ, x)
     return reduce(hadamard, kernelmatrix(κ.kernels[i], x) for i in 1:length(κ))
end
# For large matrices
A = KernelFunctions.ColVecs(rand(1000, 1000))
@btime km1($k, $A) # 5.913 s (100799 allocations) 
@btime km2($k, $A) # 5.757 s (100800 allocations)

# For small matrices
B = KernelFunctions.ColVecs(rand(10, 10))
@btime km1($k, $B) # 191.038 μs (1500 allocations: 325.70 KiB)
@btime km2($k, $B) # 192.342 μs (1501 allocations: 325.73 KiB)

So using mapreduce does not seem to give such a strong improvement

[devmotion]

Thanks for the benchmarks! Yep, I wasn't sure if this is actually useful, in particular when I noticed that in some cases (like the one you benchmarked here) we already use generators in reduce. I assume that you would see a difference mainly in cases where we actually precompute all the elements that are reduced, e.g., with map.